def __init__(
self, sess, dataset, dataset_size, logdir, ckpt, decoder_params,
model_params):
self.decoder_params = decoder_params
if self.decoder_params.token_model:
self.token_model = token_model.TokenModel(
self.decoder_params.token_model)
else:
self.token_model = token_model.TokenModel(
"speech4/conf/token_model_character_simple.pbtxt")
self.model_params = model_params
self.model_params.attention_params.type = "median"
self.model_params.attention_params.median_window_l = 10
self.model_params.attention_params.median_window_r = 100
self.model_params.tokens_len_max = 1
self.model_params.input_layer = "placeholder"
self.dataset = dataset
self.dataset_size = dataset_size
self.logdir = logdir
with tf.variable_scope("model"):
self.model = las_model.LASModel(
sess, dataset, logdir, ckpt, True, self.decoder_params.beam_width,
self.model_params)
# Graph to read 1 utterance.
tf.train.string_input_producer([dataset])
reader = tf.TFRecordReader()
filename_queue = tf.train.string_input_producer([self.dataset])
_, serialized = reader.read(filename_queue)
serialized = tf.train.batch(
[serialized], batch_size=1, num_threads=2, capacity=2)
self.features, _, self.features_len, _, _, self.text, _, _, _, _, self.uttid = s4_parse_utterance(
serialized, features_len_max=self.model_params.features_len_max,
tokens_len_max=1)
def __init__(
self, sess, dataset, dataset_size, logdir, ckpt, decoder_params,
model_params):
self.decoder_params = decoder_params
if self.decoder_params.token_model:
self.token_model = token_model.TokenModel(
self.decoder_params.token_model)
else:
self.token_model = token_model.TokenModel(
"speech4/conf/token_model_character_simple.pbtxt")
self.model_params = model_params
self.model_params.attention_params.type = "median"
self.model_params.attention_params.median_window_l = 10
self.model_params.attention_params.median_window_r = 100
self.model_params.input_layer = "decoder"
self.dataset = dataset
self.dataset_size = dataset_size
self.logdir = logdir
with tf.variable_scope("model"):
self.model = las_model.LASModel(
sess, dataset, logdir, ckpt, True, self.decoder_params.beam_width,
self.model_params)
# Graph to read 1 utterance.
tf.train.string_input_producer([dataset])
reader = tf.TFRecordReader()
filename_queue = tf.train.string_input_producer([self.dataset])
_, serialized = reader.read(filename_queue)
serialized = tf.train.batch(
[serialized], batch_size=1, num_threads=2, capacity=2)
#self.features, _, self.features_len, _, _, self.text, _, _, _, _, self.uttid = s4_parse_utterance(
# serialized, features_len_max=self.model_params.features_len_max,
# tokens_len_max=1)
self.features, self.features_fbank, self.features_len, _, self.features_weight, self.text, self.tokens, self.tokens_pinyin, self.tokens_len, self.tokens_weights, self.tokens_pinyin_weights, self.uttid = s4_parse_utterance(
serialized, features_len_max=self.model_params.features_len_max,
tokens_len_max=self.model_params.tokens_len_max + 1,
frame_stack=self.model_params.frame_stack,
frame_skip=self.model_params.frame_skip)
# Add the shape to the features.
for feature in self.features:
feature.set_shape([1, self.model_params.features_width * self.model_params.frame_stack])
for token in self.tokens:
if token:
token.set_shape([1])
def create_graph_inputs(self):
dataset_map = {}
if self.dataset == 'train_si284':
self.dataset = 'speech4/data/train_si284.tfrecords'
self.dataset_size = 37416
elif self.dataset == 'test_dev93':
self.dataset = 'speech4/data/test_dev93.tfrecords'
self.dataset_size = 503
elif self.dataset == 'test_eval92':
self.dataset = 'speech4/data/test_eval92.tfrecords'
self.dataset_size = 333
elif self.dataset == 'ptb_train':
self.dataset = 'speech4/data/ptb_train.tfrecords'
self.dataset_size = 42068
elif self.dataset == 'ptb_valid':
self.dataset = 'speech4/data/ptb_valid.tfrecords'
self.dataset_size = 3370
elif self.dataset == 'ptb_test':
self.dataset = 'speech4/data/ptb_test.tfrecords'
self.dataset_size = 3761
filename_queue = tf.train.string_input_producer([self.dataset])
reader = tf.TFRecordReader()
_, serialized = reader.read(filename_queue)
serialized = tf.train.shuffle_batch(
[serialized], batch_size=self.batch_size, num_threads=2, capacity=self.batch_size * 4 + 512,
min_after_dequeue=512, seed=1000)
# Parse the batched of serialized strings into the relevant utterance features.
self.features, self.features_len, _, self.text, self.tokens, self.tokens_len, self.tokens_weights, self.uttid = s4_parse_utterance(
serialized, features_len_max=self.features_len_max,
tokens_len_max=self.tokens_len_max + 1)
# Add the shape to the features.
for feature in self.features:
feature.set_shape([self.batch_size, self.features_width])
for token in self.tokens:
token.set_shape([self.batch_size])
